AI seat sprawl is the 2026 version of the unused SaaS licence: a Claude Team seat, a ChatGPT Business seat and a Copilot seat for the same person, bought by different managers, renewed by nobody. The overlap is not anecdotal. Ramp, which sees the card transactions, reports that "of customers that use Anthropic or OpenAI, 52% use both" (Ramp Leading Indicators, May 2026). Cledara puts the average business on its platform at "4.5 AI-related tools. Some have as many as 16" (AI Tool Sprawl, March 2026). And Zylo's 2026 SaaS Management Index finds organisations "leave an average of 36% of their SaaS licenses unused" (Zylo, January 2026).
A seat audit is the boring fix: count who produces with the model, count who only asks it for things, and stop buying producer seats for askers. This post is that audit — a worksheet, published list prices, and the one structural change that keeps the seats from coming back.
The rooms index doubles as the audit's map: which agent each group asks, how many people are in each room, how many requests they have sent, and what is running now.
"AI is quickly becoming the most expensive 'invisible worker' in the organization."
— Ben Pippenger, co-founder of Zylo, in the 2026 SaaS Management Index announcement, 29 January 2026
Who on a small team actually needs an AI seat?
Every AI seat you pay for is held by one of three kinds of person, and only one of them needs it.
- Producers are in the tool many times a day. Engineers in Claude Code or Codex, analysts living in a notebook, the designer who ships components. Their seat is a tool of the trade. Buy it, and buy the big one if they hit limits.
- Askers want outputs. A chart from this CSV, a summary of those interviews, a cleaned vendor list, a landing page in French. They open the tool twice a week, pay full price, and usually lack the files, repositories and setup the job needs anyway.
- Approvers are producers with a second job: the person whose machine and repositories the work touches, who should see what runs and be able to stop it.
Per-seat plans price all three identically. That is the sprawl, and it grows quietly rather than loudly. Zylo's index records expense-based SaaS spend up "267% year over year, with ChatGPT now the most expensed application" — a polite way of saying those seats are bought on personal cards, outside procurement, by people who will never tell you they stopped using them.
How do you run an AI seat audit in one afternoon?
- List every AI seat, across vendors, with the person and the monthly price. Include the ones expensed on personal cards. Zylo puts 81% of SaaS spend in the hands of business units rather than IT, so the list you already have is not the list.
- Pull usage. Every major vendor hands an admin this. Anthropic's Analytics API returns "per-user daily metrics" across chat and Claude Code, and, at the organisation level, "daily, weekly, and monthly active users, seat counts, and pending invites" (Anthropic docs) — seats against actives is the whole audit in one row. GitHub lets you "sort the list of users by when they last used GitHub Copilot" and download an activity report as CSV (GitHub docs). OpenAI's own admin vocabulary is the sharpest of the three: seats purchased, seats enabled, seats activated, and weekly active users (OpenAI Academy). The gap between purchased and activated is the money.
- Classify each person as producer, asker or ghost. Fewer than eight active days in the last sixty means asker or ghost. Be honest about founders. One caveat before you cancel on a number: GitHub notes that last-activity data "may require up to 24 hours to update and depends on telemetry being enabled in user IDE settings," so a quiet seat is not always a dead seat. Ask the person.
- Act. Ghosts: cancel. Askers: downgrade to free tiers, and route their real asks through a producer's agent (below). Producers: keep, and upgrade the ones who hit limits.
- Set the rule so the sprawl does not come back: a seat unused for sixty days is downgraded automatically, and a new seat requires a producer justification.
What does the audit look like on a real team?
Six people, nine seats, one afternoon.
Five of those nine seats have a published price. Anthropic lists a Claude Team seat at $25 per seat per month billed monthly, $20 billed annually; Claude Pro at $20 monthly, $17 annually; and Claude Max from $100 a month. GitHub lists Copilot Business at "$19 USD per granted seat per month" — GitHub bills the granted seat, not the use of it, which is the entire problem in five words.
At those monthly rates the priced half of the table is $194 a month before the audit (three Claude Team seats at $25, one Copilot Business seat at $19, Noah's Max at $100) and $120 after (Noah's Max, plus Sam on Pro at $20). That is $74 a month recovered, or 38%, before a single ChatGPT seat is counted, and four of the nine seats in the table are ChatGPT Business.
We are not printing a ChatGPT price here. OpenAI states only that "paid plans (Go, Plus, Business, and Enterprise) are priced per user per month" and that "Business plans are available starting at 2 users"; read the current figure on OpenAI's own pricing page rather than trusting a number in a blog post, including this one.
The better number is not the dollars anyway. It is three seats used every day instead of nine used twice a week.
How do you keep the seats from coming back?
Audits fail when askers still need outputs and have nowhere to get them. They buy the seat back in a month, on a card, quietly — which is exactly how it got expensed the first time.
The structural fix is to give askers a place to ask. In muti, a producer connects their machine, creates an agent from the Claude Code or Codex already installed there, and invites the askers into a room, or gives the agent a Slack bot so they can DM it or mention it in a channel. The asker describes the job and attaches the file. The work runs on the producer's machine, where the data, repositories and tools already are, under the producer's own subscription. Everyone watches the same timeline. Risky steps pause for the producer; routine ones become standing rules after the first approval. Each request carries an id, a requester, an approver and a cost.
A request in a room: the asker, the id, the tool calls, the finished chart, and the cost of that one request, visible to everyone in the room.
Three things follow for the audit:
- Askers lose nothing they used. They get faster answers than a seat gave them, because the agent runs where the files are.
- Usage becomes visible per request, not per seat. When one asker's errands start to crowd a producer's plan, the request costs say so, and that is the signal to add a producer seat, with a reason.
- The sixty-day rule has teeth. A downgraded asker still has the room. There is no reason to buy the seat back.
When should you still buy per-seat plans?
Per-seat plans are the right answer for producers, for teams where most people are producers, and for organisations that need contractual data terms for every user; buy Team, Business or Enterprise for those. Anthropic sells Team "for teams of 2 to 150," so the small-team case is explicitly covered — the question is only how many of your people are producers.
A room does not share anyone's account: the producer signs in, the producer's agent runs, and other people's requests are work the producer chose to accept and can stop. Vendor terms prohibit sharing credentials and reselling access, and none of that happens here. What the room removes is the seat bought for someone who wanted a chart.
Frequently asked questions
Won't the producer's own usage limit suffer?
Their plan now carries other people's errands, so yes, some. Cost per request is next to each request; if the errands become substantial, buy a second producer seat. The audit is not about zero seats, it is about the right number.
How do we know what askers did with the agent?
Every request has an id, a timeline and an approver, visible to the room. It is a better audit trail than per-seat chat history you cannot see. Vendor consoles are deliberately coarser here: OpenAI notes that its own Task Insights "is designed for aggregate analysis only and does not expose individual prompts or conversations."
Does the producer have to be online?
Requests wait in the agent's inbox and run when the machine is back; approvals wait 24 hours. Routine, already-allowed work runs without them. See the agent inbox.
Which runtimes work?
Claude Code, Codex, Cursor, OpenCode and more; see the Claude version and the Codex version of this arithmetic.
How often should we re-run the audit?
Quarterly, and once more whenever a vendor changes its pricing model. Flexera's 2026 State of ITAM report found 59% of organisations reporting increased wasted AI spend while "only 31% of organizations have visibility into AI software" — the waste is a visibility problem, and visibility decays.
Sources
Prices and product behaviour were read on the vendors' own pages on 14 September 2026. Check them before you act; they change.
- Anthropic, Claude pricing — Team, Pro and Max seat prices, and the 2-to-150 team range.
- Anthropic, Analytics API and usage analytics for Team and Enterprise — per-user activity, seat counts and per-user spend.
- GitHub, plans for GitHub Copilot and reviewing user activity data — per-granted-seat pricing, last-used sorting, CSV export.
- OpenAI, ChatGPT pricing and Enterprise workspace analytics guide — per-user billing, the 2-user minimum, and the purchased / enabled / activated seat funnel.
- Zylo, 2026 SaaS Management Index, 29 January 2026 — 36% of licences unused, expense-based spend up 267%, 81% of spend outside IT.
- Ramp, The AI race has a new frontrunner, but no clear winner, 19 May 2026 — 52% of Anthropic or OpenAI customers use both.
- Cledara, AI Tool Sprawl, 27 March 2026 — 4.5 AI tools per business on average, up to 16.
- Flexera, State of ITAM 2026, 24 June 2026 — 59% report increased wasted AI spend; 31% have visibility into AI software.
Audit the seats. Keep the producers. Give the askers a room.
Published · Updated